Process node text monitoring method and system based on AI natural language processing

Abstract

The embodiment of the invention provides a process node text monitoring method and system based on AI natural language processing, and relates to the technical field of business process management.The method comprises the steps that firstly, text interaction data and corresponding process progress information in the execution process of each process node are captured, and then correlation modeling processing is conducted on the text interaction data and the corresponding process progress information; obtaining a corresponding relation and a task stage matching degree, calling a pre-training model to perform exception tracing processing, generating a text interaction exception tracing result, and determining exception triggering factors and corresponding process task links according to the text interaction exception tracing result; and finally, a process node optimization instruction containing a node interaction rule adjustment suggestion is generated based on the information, and is sent to a process control terminal, so that process node text interaction can be accurately monitored, abnormity can be timely found and positioned, dynamic optimization of the process node is realized, and the intelligent level of business process management is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of business process management, in particular to a process node text monitoring method and system based on AI natural language processing. BACKGROUND

[0002] In today's complex business process management scenarios, efficient and accurate execution of process nodes plays a key role in the smooth operation of the overall business process. With the expansion of business scale and the increase of process complexity, a large amount of text interaction data is generated during the execution of process nodes, which contains rich business information and is crucial for monitoring the execution status of process nodes.

[0003] Traditional process node monitoring methods mainly focus on monitoring hard indicators such as process progress and time, often ignoring in-depth analysis of text interaction data. Some methods may record text information, but only store it simply, lacking in-depth mining of the association between text content and process task stages. This makes it difficult to quickly and accurately locate the problem source from text interaction data when the process is abnormal, and it is impossible to determine whether the abnormality is caused by text interaction content not meeting requirements or not matching the current task stage. Moreover, traditional methods lack effective optimization mechanisms after discovering abnormalities, and cannot adjust and optimize process nodes in a timely manner according to abnormal conditions, thereby affecting the overall efficiency and stability of business processes. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a process node text monitoring method and system based on AI natural language processing.

[0005] According to the first aspect of the present application, a process node text monitoring method based on AI natural language processing is provided, which comprises: capturing text interaction data generated by each process node during process execution and corresponding process progress information, the text interaction data containing text content sent and received by each process node, and the process progress information containing the task stage currently executed by each process node; associating and modeling the text interaction data and the process progress information to obtain the corresponding relationship between text interaction and task stage of each process node and the task stage matching degree corresponding to text interaction; calling a pre-trained process node text monitoring model to perform abnormality tracing processing on the corresponding relationship and the task stage matching degree to generate text interaction abnormality tracing results of each process node; determining the trigger factors of text interaction abnormalities of each process node and the process task links corresponding to abnormal text interaction according to the text interaction abnormality tracing results; generate a process node optimization instruction containing a node interaction rule adjustment suggestion based on the trigger factor and the process task link, and send the process node optimization instruction to a process control terminal.

[0006] According to a second aspect of the present application, a process node text monitoring system based on AI natural language processing is provided, which comprises a processor and a readable storage medium, and the readable storage medium stores a program which, when executed by the processor, implements the aforementioned process node text monitoring method based on AI natural language processing.

[0007] According to any one of the above aspects, by capturing text interaction data and corresponding process progress information in the process node execution process, the text interaction data and the process progress information are associated and modeled, the correspondence between the text interaction of each process node and the task stage and the matching degree are accurately obtained, the association between the text interaction and the business process is more clear and explicit, the pre-trained process node text monitoring model is called for abnormal trace processing, the powerful natural language processing capability of AI is utilized to deeply analyze the text interaction data, accurately generate the text interaction abnormal trace result of each process node, greatly improve the accuracy and efficiency of abnormal positioning, the trigger factor and the corresponding process task link are determined according to the abnormal trace result, the cause of the problem is fundamentally found out, the process node optimization instruction containing the node interaction rule adjustment suggestion is generated based on the trigger factor and the process task link, and is sent to the process control terminal, the dynamic optimization and adjustment of the process node are realized, the problems existing in the process are solved in time, and the stability, accuracy and execution efficiency of the business process are improved. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 A flowchart of the process node text monitoring method based on AI natural language processing provided by the embodiments of the present application is shown. Figure 2 A component structure schematic diagram of the process node text monitoring system based on AI natural language processing provided by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0009] Figure 1 A flowchart of the process node text monitoring method based on AI natural language processing provided by the embodiments of the present application is shown. It should be understood that in other embodiments, the order of some steps of the process node text monitoring method based on AI natural language processing can be exchanged according to actual needs, or some steps can be omitted or deleted. The detailed steps of the process node text monitoring method based on AI natural language processing are as follows.

[0010] Step S110: Capture the text interaction data generated by each process node during the process execution and the corresponding process progress information. The text interaction data includes the text content sent and received by each process node. The process progress information includes the current task stage of each process node.

[0011] In the process monitoring scenario of a software development project, the process nodes include requirement analysis nodes, design nodes, development nodes, testing nodes, and deployment nodes, etc. Each node can generate a large amount of text interaction data when performing the corresponding task. For example, the requirement analysis node exchanges text messages with the customer through a project management platform, which includes the text content sent by the customer, such as functional requirement description, modification suggestions for the requirement document, etc., and the text content received by the requirement analysis node, such as requirement understanding confirmation information, requirement collection progress information, etc.

[0012] Meanwhile, each process node updates its process progress information in real time. The task stages of the requirement analysis node may include requirement collection stage, requirement sorting stage, and requirement document writing stage. The task stages of the development node may include module development stage, interface development stage, and unit testing stage, etc. These task stage information is recorded in real time as a core component of the process progress information. Through the data flow capture components deployed in the project management platform and the node terminals, the above-mentioned text interaction data and process progress information can be captured in real time and stored in the data processing center.

[0013] Step S120: Perform associated modeling processing on the text interaction data and the process progress information to obtain the corresponding relationship between the text interaction of each process node and the task stage and the matching degree of the text interaction corresponding to the task stage.

[0014] After obtaining the text interaction data and the process progress information, the association between the two needs to be established to clarify in which task stage the text interaction occurs and the matching degree of the text interaction and the corresponding task stage. This process needs to be realized through a series of semantic analysis and feature matching operations to ensure that the text interaction data can be accurately corresponded to the task stage it belongs to and the matching level between the two is quantified.

[0015] Step S121: Perform semantic analysis processing on the text content in the text interaction data to extract semantic elements representing the task execution state in the text content. The semantic elements include task execution progress description, task requirement description, and task result feedback.

[0016] For the captured text interaction data, the semantic analysis of the text content is performed first. Taking the text content of the development node in the module development stage as an example, if the text content sent by the developer is “User login module has completed 70% of code writing, and is expected to complete unit test case design tomorrow”, after the semantic analysis tool performs word segmentation, part-of-speech tagging and dependency syntax analysis on the text, the task execution progress description “User login module has completed 70% of code writing” and the task result feedback related content “Expected to complete unit test case design tomorrow” can be extracted.

[0017] For the customer text content “Hope to add mobile verification code function during user registration to ensure account security” received by the requirement analysis node, the semantic analysis extracts the task requirement description “Add mobile verification code function during user registration to ensure account security”. These extracted semantic elements will serve as the basis data for subsequent association modeling.

[0018] Step S1211: The text content in the text interaction data is divided into multiple sentence units according to the sentence division rule.

[0019] The sentence division rule is set based on punctuation marks and semantic pauses. When a punctuation mark indicating the end of a sentence such as a period, question mark or exclamation mark is encountered, the text content is split into an independent sentence unit. For example, the text “Today has completed the requirement collection work, a total of 25 functional requirements have been collected. Start demand analysis tomorrow to form a preliminary demand list” sent by the requirement analysis node is split into two sentence units “Today has completed the requirement collection work, a total of 25 functional requirements have been collected” and “Start demand analysis tomorrow to form a preliminary demand list” according to the rule.

[0020] Step S1212: A pre-trained semantic understanding model is called to convert each sentence unit into a sentence semantic vector that can represent its semantic connotation.

[0021] The pre-trained semantic understanding model uses a language model based on the Transformer architecture. This semantic understanding model has been trained on a large amount of text corpus and can understand the semantic connotation of the text. Each split sentence unit is input into the model, which analyzes the context relationship of the words in the sentence unit and outputs a multi-dimensional sentence semantic vector. Each sentence semantic vector is composed of multiple numerical values, and different numerical value combinations correspond to different semantic connotations. For example, the sentence semantic vector corresponding to “complete requirement collection work” is significantly different from the sentence semantic vector corresponding to “not start requirement collection” in numerical distribution.

[0022] Step S1213: retrieve a preset task status semantic dictionary, the task status semantic dictionary comprising feature words representing task execution states and corresponding semantic vectors, the feature words comprising task execution in progress, task to be confirmed, task completed, and task to be adjusted.

[0023] The task status semantic dictionary is pre-constructed, wherein each feature word is processed by a semantic understanding model to generate a corresponding semantic vector and stored. For example, the semantic vector corresponding to "task execution in progress" records the feature distribution of the word in the semantic space, and the semantic vector corresponding to "task completed" embodies the semantic features of the completion state.

[0024] Step S1214: calculate the similarity between the sentence semantic vector of each sentence unit and the semantic vector of each feature word in the task status semantic dictionary.

[0025] The similarity calculation uses the cosine similarity calculation method to measure the similarity between two vectors by calculating the cosine value of the included angle between them. Specifically, the sentence semantic vector of each sentence unit and the semantic vector of each feature word in the task status semantic dictionary are calculated respectively to obtain multiple similarity values. For example, the similarity value calculated between the sentence semantic vector of the sentence unit "user login module has completed 70% of the code writing" and the semantic vector of the "task execution in progress" feature word is relatively high, while the similarity value calculated between the sentence semantic vector of the sentence unit and the semantic vector of the "task completed" feature word is relatively low.

[0026] Step S1215: filter out the semantic information associated with the semantic vector corresponding to the feature word with the highest similarity, and determine the semantic information as the semantic element representing the task execution state in the sentence unit.

[0027] According to the similarity value calculated in the previous step, the feature word with the largest value is selected. The semantic information corresponding to the feature word is the semantic element representing the task execution state in the sentence unit. For example, if the sentence semantic vector of the sentence unit has the highest similarity with the semantic vector of the "task execution in progress" feature word, the semantic information related to "task execution in progress" is determined as the semantic element of the sentence unit. For the sentence unit containing specific progress description, the specific progress information is also supplemented according to the sentence content to form a complete semantic element.

[0028] Step S1216: integrate the semantic elements corresponding to all sentence units in each text content to obtain the semantic elements representing the task execution state in the text content.

[0029] The semantic elements determined by all the split sentence units of the same text content are integrated. If the semantic elements of different sentence units have a correlation or supplementary relationship, they are combined into a complete semantic element set. For example, two sentence units of certain text content extract the semantic elements "task execution" and "expected completion tomorrow" respectively, and after integration, the semantic element of the text content is "task execution, expected completion tomorrow".

[0030] Step S122: Feature identification processing is performed on the task stages in the process progress information, and a unique stage feature identifier is set for each task stage. The stage feature identifier includes the task stage name, the task stage execution requirement, and the task stage output standard.

[0031] Feature identification processing is performed on each task stage in the process progress information. Taking a test node as an example, its task stages include a unit test stage, an integration test stage, and a system test stage. In the stage feature identifier set for the unit test stage, the task stage name is "unit test stage", the task stage execution requirement is "independent testing of each developed module, covering all function points and boundary conditions in the module", and the task stage output standard is "generating a unit test report with a defect pass rate of not less than 95%".

[0032] For the module development stage of a development node, the task stage name of the stage feature identifier is "module development stage", the execution requirement is "code writing according to the design document, following the coding specification, and local debugging every 50 lines of code", and the output standard is "completion of module code writing, submission to the code repository, and passing static code checking". The stage feature identifier of each task stage is stored in the feature identifier library to ensure uniqueness and accuracy.

[0033] Step S123: The extracted semantic elements are subjected to semantic matching processing with the stage feature identifiers of the task stages, to determine the task stage corresponding to each text interaction data, and form the corresponding relationship between the text interaction and the task stage of each process node.

[0034] The semantic elements extracted from the text interaction data are subjected to semantic matching with the stage feature identifiers of the task stages in the feature identifier library. The semantic matching process is realized by calculating the semantic similarity of the semantic elements and the task stage execution requirements, output standards, etc. in the stage feature identifier. For example, the semantic element of a certain text interaction data is "independent testing of the user management module is in progress, and 80% of the test cases have been executed", which has a high semantic similarity with the stage feature identifier of the unit test stage of the test node in terms of execution requirements, so it is determined that the task stage corresponding to the text interaction data is the unit test stage.

[0035] By the above matching mode, the corresponding task stage is found for each text interaction data, and then the corresponding relationship such as "development node module development stage-text interaction data A" and "test node unit test stage-text interaction data B" is formed, and recorded in the association relationship table.

[0036] Step S124: Calculate the semantic similarity between the semantic element of each text interaction data and the stage feature identifier of the corresponding task stage, and take the semantic similarity as the task stage matching degree corresponding to the text interaction.

[0037] After determining the task stage corresponding to the text interaction data, the semantic similarity between the semantic element and the stage feature identifier of the task stage is calculated. When calculating, the semantic association degree of the semantic element and the task stage name, execution requirement and output standard in the stage feature identifier are comprehensively considered. For example, the semantic element of the text interaction data is "complete user login module code writing, meet the coding specification, and pass the static code check", and the corresponding task stage is the module development stage of the development node.

[0038] The semantic similarity between the semantic element and the module development stage stage feature identifier is calculated. If the semantic element completely covers the output standard and part of the execution requirement of the stage feature identifier, the obtained semantic similarity value is higher, and the value is taken as the task stage matching degree corresponding to the text interaction. The matching degree is valued between 0 and 1, and the value closer to 1 indicates that the matching degree of the text interaction and the task stage is higher.

[0039] Step S125: Perform node identification binding processing on the corresponding relationship and the task stage matching degree, so that each corresponding relationship and task stage matching degree corresponds to the flow node identifier generating the text interaction data, and obtain the corresponding relationship between the text interaction and the task stage of each flow node and the task stage matching degree corresponding to the text interaction.

[0040] Each flow node is assigned a unique node identifier, such as a requirement analysis node identifier J001, a development node identifier J002, etc. The previously formed corresponding relationship between the text interaction and the task stage and the calculated task stage matching degree are bound with the flow node identifier generating the text interaction data. For example, the corresponding relationship and matching degree of a certain text interaction data of the J002 development node in the module development stage are bound with the J002 identifier.

[0041] The bound information is stored in a structured data table, which includes node identifier, text interaction data ID, corresponding task stage, task stage matching degree and other fields, facilitating subsequent model calling and data query.

[0042] Step S130: calling the pre-trained process node text monitoring model to perform abnormality tracing processing on the correspondence and the task stage matching degree, and generating text interaction abnormality tracing results of each process node.

[0043] After obtaining the correspondence between the text interaction of each process node and the task stage and the task stage matching degree, these data are input into the pre-trained process node text monitoring model. The model identifies possible abnormal patterns in the data through deep analysis, traces the source and path of the abnormality, and finally generates abnormality tracing results containing abnormal characteristics, abnormal transmission paths, and other information.

[0044] Step S131: inputting the correspondence and the task stage matching degree into the feature association module of the process node text monitoring model, performing structured coding processing on the correspondence to obtain a relationship feature vector, and performing numerical feature conversion processing on the task stage matching degree to obtain a matching degree feature vector.

[0045] The feature association module of the process node text monitoring model receives the input correspondence and task stage matching degree. For the correspondence, structured coding is used to convert text interaction data ID, corresponding task stage identifier, node identifier, and other information into a fixed-length relationship feature vector. Each dimension in the relationship feature vector corresponds to the encoding value of a structured information, for example, node identifier J002 corresponds to a specific dimension value in the vector.

[0046] For the task stage matching degree, numerical feature conversion processing is performed to map the matching degree value to a specific numerical interval in proportion, forming a matching degree feature vector. The matching degree feature vector is a single-dimensional vector, and its numerical value corresponds to the numerical value of the task stage matching degree, ensuring the stability of the numerical feature in model processing.

[0047] Step S132: in the feature association module of the process node text monitoring model, performing feature fusion processing on the relationship feature vector and the matching degree feature vector to generate a fusion feature vector, and the feature fusion processing is performed by feature connection processing to make the fusion feature vector contain correspondence information and matching degree information.

[0048] The feature fusion processing in the feature association module uses feature connection to sequentially splice the relationship feature vector and the matching degree feature vector together to form a fusion feature vector. For example, the relationship feature vector is a vector containing 100 dimensions, and the matching degree feature vector is a vector containing 1 dimension, and after splicing, a fusion feature vector containing 101 dimensions is formed.

[0049] The first 100 dimensions of the fusion feature vector retain the information of the relationship feature vector, and the 101st dimension retains the information of the matching degree feature vector, so that the fusion feature vector carries information of both the correspondence and the matching degree.

[0050] Step S133: inputting the fusion feature vector into an anomaly tracing module of the process node text monitoring model, the anomaly tracing module of the process node text monitoring model comprising a plurality of parallel tracing sub-modules, each tracing sub-module corresponding to a text interaction anomaly type.

[0051] The anomaly tracing module comprises a plurality of parallel tracing sub-modules, each corresponding to a different text interaction anomaly type, such as expression deviation anomaly, stage misplacement anomaly, rule violation anomaly, etc. The fusion feature vector is simultaneously input into all the tracing sub-modules, each sub-module focusing on identifying features related to the anomaly type it corresponds to. For example, the expression deviation anomaly tracing sub-module focuses on analyzing the feature dimensions in the fusion feature vector related to text semantic expression, and the stage misplacement anomaly tracing sub-module focuses on the feature dimensions of the correspondence between text interaction and task stage.

[0052] Step S134: each tracing sub-module performs anomaly feature extraction processing on the fusion feature vector to identify feature elements in the fusion feature vector related to the corresponding anomaly type, the feature elements including text interaction semantic features corresponding to the anomaly and task stage features corresponding to the anomaly.

[0053] Each tracing sub-module internally comprises a feature extraction layer that processes the input fusion feature vector through a multi-layer neural network. Taking the expression deviation anomaly tracing sub-module as an example, its feature extraction layer focuses on the dimensions in the fusion feature vector related to text semantics and highlights the features of semantic expression anomalies through weight adjustment.

[0054] During processing, dimensions in the fusion feature vector with significantly different numerical distributions from normal features are identified, and the features corresponding to these dimensions are the text interaction semantic features corresponding to the anomaly. At the same time, in combination with the task stage related feature dimensions, the task stage features corresponding to the anomaly are extracted, such as features with inconsistent task stage identifiers.

[0055] Step S1341: in each tracing sub-module, retrieve the anomaly feature template associated with the corresponding anomaly type, the anomaly feature template being a typical feature distribution form of the fusion feature vector for this anomaly type.

[0056] Each traceability submodule stores an abnormal feature template associated with the abnormal type corresponding thereto. The abnormal feature template is summarized by analyzing a large amount of historical abnormal data, and embodies the typical feature distribution form of the fusion feature vector under the abnormal type. For example, in the abnormal feature template of the stage misplacement abnormality, there is a specific difference pattern between the dimension value corresponding to the task stage identifier in the fusion feature vector and the value in the normal case, and there is a mismatch feature distribution between the task stage field corresponding to the text interaction data and the actual task stage field.

[0057] Step S1342: performing feature comparison processing on the fusion feature vector and the abnormal feature template to identify a feature region in the fusion feature vector that is consistent with the feature distribution of the abnormal feature template.

[0058] The fusion feature vector and the abnormal feature template are compared dimension by dimension, and the numerical difference between the two in each dimension is calculated. When the numerical difference of multiple consecutive dimensions conforms to the feature distribution pattern of the abnormal feature template, the region of these dimensions is determined as a feature region that is consistent with the feature distribution of the abnormal feature template. For example, in the stage misplacement abnormality traceability submodule, if the numerical values of multiple dimensions corresponding to the task stage in the fusion feature vector are consistent with the numerical distribution of the region in the abnormal feature template, the region is identified as a feature region.

[0059] Step S1343: extracting feature elements in the feature region, including the numerical value of the feature element, the position of the feature element in the fusion feature vector, and the text interaction and task stage association information corresponding to the feature element.

[0060] Specific feature elements are extracted from the identified feature region, and the numerical value of each feature element is recorded to determine its position index in the fusion feature vector. At the same time, according to the dimension information corresponding to the feature element, the corresponding text interaction and task stage association information is deduced, such as the text interaction data ID corresponding to the feature element, the task stage identifier, etc. These information collectively constitute the complete attributes of the extracted feature elements.

[0061] Step S1344: performing semantic association verification processing on the extracted feature elements to confirm the semantic association between the text interaction semantics corresponding to the feature elements and the abnormal type, and the semantic association is determined by the similarity between the semantic vector corresponding to the feature element and the semantic vector of the abnormal type.

[0062] The text interaction semantics corresponding to the extracted feature element is converted into a semantic vector, and an abnormal type semantic vector corresponding to the abnormal type of the traceability submodule is called. The similarity of the two semantic vectors is calculated. If the similarity value is greater than a preset threshold, it indicates that the text interaction semantics corresponding to the feature element has semantic relevance with the abnormal type; otherwise, it does not have relevance. Through the above verification, it is ensured that the extracted feature element is indeed related to the abnormal type.

[0063] Step S1345: The feature element with semantic relevance meeting the preset requirement is determined as a feature element related to the corresponding abnormal type.

[0064] According to the result of semantic association verification, the feature elements with similarity values greater than the preset threshold are selected, and these feature elements are formally determined as feature elements related to the corresponding abnormal type. These feature elements will serve as an important basis for subsequent tracing of the abnormal generation path, ensuring the accuracy of abnormal traceability.

[0065] Step S135: The path of abnormal generation is traced according to the extracted feature elements, including the sending node, receiving node and transmission order of abnormal text interaction in the process.

[0066] Based on the determined abnormal feature elements, the path of abnormal generation is traced in combination with the text interaction corresponding to the feature elements and the task stage association information. By querying the association relationship table and the text interaction record, the sending node identifier and the receiving node identifier of the abnormal text interaction are determined. For example, the text interaction data corresponding to the feature element is sent by the J002 development node, and the receiving node is the J003 test node, so the sending node is J002 and the receiving node is J003.

[0067] At the same time, according to the timestamp information of the text interaction data, the transmission order of the abnormal text interaction in the process is arranged. For example, if the J002 development node sends an abnormal text interaction data to the J003 test node in the module development stage, the timestamp is T1; then the J003 test node sends related text interaction data based on the abnormal text interaction to the J004 deployment node, the timestamp is T2, and T2 is later than T1, so the transmission order of the abnormal text interaction in the process is J002—J003—J004. Through the above method, the entire path of abnormal generation and transmission is completely traced, and the flow of the abnormal in each node is determined.

[0068] Step S136: The abnormal feature elements and abnormal transmission paths output by each traceability submodule are integrated to generate text interaction abnormal traceability results for each process node.

[0069] Collect the abnormal feature elements and corresponding abnormal transmission paths output by all traceability sub-modules, and integrate these information. For each process node, sort out the abnormal content related to it, including the specific description of the abnormal feature elements involved in the node, the type of the abnormality, and the transmission of the abnormality at the node. For example, the abnormal traceability result of the J002 development node can include the semantic features of the abnormal text interaction sent by it, the corresponding stage misalignment abnormal type, and the path information of the abnormality transmitted to the J003 test node. The integrated result is presented in the form of a structured report, forming the text interaction abnormal traceability result of each process node.

[0070] Step S140: Determine the trigger factors of the text interaction abnormality of each process node and the process task link corresponding to the abnormal text interaction according to the text interaction abnormal traceability result.

[0071] After obtaining the text interaction abnormal traceability result, analyze the abnormal features, transmission paths, and other information in the result in depth, find out the root cause of the abnormality, i.e., the trigger factor, and identify the specific process task link to which the abnormal text interaction belongs.

[0072] Step S141: Analyze the abnormal feature elements in the text interaction abnormal traceability result, identify the semantic information representing the abnormal inducement in the abnormal feature elements, and the semantic information includes text expression deviation, task stage cognitive deviation, and interaction rule understanding deviation.

[0073] Perform semantic analysis on the abnormal feature elements in the text interaction abnormal traceability result, and extract the semantic information that can reflect the abnormal inducement. For example, the text interaction content corresponding to a certain abnormal feature element is "module development has been completed, integration testing can be performed", but actually the module has only completed code writing and has not passed static code checking. Through analysis, it is found that the semantic information is text expression deviation, i.e., the text content does not match the actual task status. If the abnormal feature element shows that the task stage corresponding to the text interaction is the integration testing stage, while the actual process is in the unit testing stage, the semantic information is task stage cognitive deviation.

[0074] Step S142: Retrieve the text interaction historical data of each process node, analyze the text interaction rules of each process node before the abnormality occurs, and the text interaction rules include text interaction frequency, text interaction content type, and task stage distribution corresponding to text interaction.

[0075] Retrieve the text interaction historical data of each process node before the abnormality occurs from the data storage center. These data cover all text content sent and received by each node in the past period of time and the corresponding task stage. Through the statistics and analysis of these historical data, the text interaction rules under normal circumstances are summarized.

[0076] Step S1421: Obtain text interaction history data of each process node in a preset time period before the abnormality occurs, the text interaction history data containing all text content sent and received by each process node in the time period and a corresponding task stage.

[0077] The preset time period can be set according to the task cycle of the process node, for example, set to 30 days before the abnormality occurs. In this time period, the text interaction history data of all process nodes such as the J001 requirement analysis node and the J002 development node is collected, including the specific text information of each text content, the sending and receiving time, and the detailed content of the corresponding task stage identifier, to ensure the completeness and accuracy of the data.

[0078] Step S1422: Classify the text content in the text interaction history data, and divide the text content into task instruction type, task feedback type and task consultation type according to the semantic type of the text content.

[0079] According to the classification of the semantic type of the text content, the task instruction type text content mainly contains the task arrangement and operation requirements of the superior node to the subordinate node, such as “please complete the unit test of the user module within this week”; the task feedback type text content includes the report on the task execution situation and the result explanation, such as “the unit test of the user module has been completed, and 3 defects have been found”; and the task consultation type text content is the inquiry between nodes about the task details and technical problems, such as “what is the interface call specification of the integration test”. Through classification, the distribution of different types of text content is clear.

[0080] Step S1423: Count the number of occurrences of each type of text content in different task stages, calculate the proportion of each type of text content in each task stage, and obtain the distribution of text interaction content types in each task stage.

[0081] The classified text content is counted, and the number of occurrences of the task instruction type, the task feedback type and the task consultation type text in each task stage is recorded. For example, in the module development stage of the development node, it is counted that the task instruction type text appears 20 times, the task feedback type text appears 35 times, and the task consultation type text appears 15 times, and the total number of text occurrences in this stage is 70 times. The proportion of the task instruction type text in this stage is 20 / 70, the proportion of the task feedback type is 35 / 70, and the proportion of the task consultation type is 15 / 70, so as to obtain the distribution of text interaction content types in each task stage.

[0082] Step S1424: Count the total number of text content sent and received by each process node in a preset time interval, calculate the number of text interactions per unit time, and obtain the text interaction frequency of each process node.

[0083] The preset time interval can be set as every day, and the total number of times of sending and receiving text content of each process node per day is counted. For example, the J002 development node has an average of 45 times of sending and receiving text content per day in the past 30 days, and the text interaction frequency of the node is 45 times per day. In this way, the average text interaction frequency of each process node is obtained, reflecting the text interaction activity level of the node.

[0084] Step S1425: integrate the distribution and the text interaction frequency to form the text interaction rule of each process node before the anomaly occurs.

[0085] The distribution of text interaction content types in each task stage and the text interaction frequency of each process node are integrated to form a complete text interaction rule. For example, the text interaction rule of the J002 development node is: in the module development stage, the task feedback text accounts for 50%, the task instruction text accounts for 28.6%, the task consultation text accounts for 21.4%, and the text interaction frequency is 45 times per day; in the interface development stage, the task feedback text accounts for 40%, the task instruction text accounts for 30%, the task consultation text accounts for 30%, and the text interaction frequency is 50 times per day, etc.

[0086] Step S143: comparing the text interaction data when the anomaly occurs with the text interaction rule to identify the difference between the two, the difference point including text interaction frequency change, text interaction content type change, and text interaction corresponding task stage deviation.

[0087] The text interaction data when the anomaly occurs is compared with the previously summarized text interaction rule item by item. If the text interaction frequency of the J002 development node becomes 20 times per day when the anomaly occurs, which is obviously different from the rule of 45 times per day, the difference point is the text interaction frequency change. If the task consultation text accounts for 60% in the module development stage, which is inconsistent with the rule of 21.4%, the difference point is the text interaction content type change. If the text interaction corresponds to the integration test stage, but the actual process is in the unit test stage, the difference point is the text interaction corresponding task stage deviation.

[0088] Step S144: determining the trigger factor of the text interaction anomaly according to the semantic information and the difference point, the trigger factor including unclear text expression, task stage cognitive error, and interaction rule execution deviation.

[0089] In combination with the parsed semantic information and the identified difference points, the trigger factor of the text interaction abnormality is comprehensively judged. If the semantic information is text expression deviation, and the difference points show that the text content type changes, it can be determined that the trigger factor is unclear text expression. If the semantic information is task stage cognitive deviation, and there are difference points of task stage offset, the trigger factor is task stage cognitive error. If the interaction rules stipulate that in the unit test stage, task feedback type text should be mainly sent, and a large number of task instruction type text actually appears, the difference points are content type changes, and the semantic information involves interaction rule understanding deviation, then the trigger factor is interaction rule execution deviation.

[0090] Step S145: The link division information of the process task is called, and the process task link to which the abnormal text interaction belongs is determined according to the task stage corresponding to the abnormal text interaction, to obtain the process task link corresponding to the abnormal text interaction.

[0091] The link division information of the process task is called from the process management database, which clearly shows the process task link to which each task stage belongs. For example, the process task link division of a software development project is divided into a requirement analysis link, a design link, a development link, a test link and a deployment link, wherein the development link includes a module development stage, an interface development stage and a unit test stage. If the task stage corresponding to the abnormal text interaction is the module development stage, the process task link to which it belongs is determined according to the link division information, that is, the process task link corresponding to the abnormal text interaction is obtained.

[0092] Step S150: Based on the trigger factor and the process task link, the process node optimization instruction containing the node interaction rule adjustment suggestion is generated, and the process node optimization instruction is sent to the process control terminal.

[0093] After the trigger factor and the process task link are determined, node interaction rule adjustment suggestions are made for specific problems, and these suggestions and abnormal related information are integrated into process node optimization instructions, which are sent to the process control terminal to realize the optimization and adjustment of the process node interaction.

[0094] Step S151: The preset abnormal trigger factor coping strategy library is called, and the abnormal trigger factor coping strategy library contains the interaction rule adjustment direction and adjustment implementation points corresponding to each trigger factor.

[0095] The abnormal trigger factor coping strategy library is pre-established and stores standard coping strategies corresponding to various trigger factors. For example, for the trigger factor of unclear text expression, the interactive rule adjustment direction is to standardize the text expression format, and the adjustment implementation points include clearly defining the element composition of the text content, using standardized terminology, etc. For the task stage cognitive error, the adjustment direction is to strengthen the management of task stage identification, and the implementation points include clearly marking the task stage in the text interaction and regularly conducting stage cognitive training, etc.

[0096] Step S152: According to the trigger factor, the corresponding interactive rule adjustment direction and adjustment implementation points are matched from the abnormal trigger factor coping strategy library.

[0097] According to the determined text interaction abnormal trigger factor, a search and match are performed in the abnormal trigger factor coping strategy library. If the trigger factor is unclear text expression, the corresponding interactive rule adjustment direction "standardize the text expression format" and the adjustment implementation points "clearly define the three elements of task name, execution status, and time node that the text content should contain, and use standardized terminology in the project terminology table" are matched.

[0098] Step S153: The matched interactive rule adjustment direction and adjustment implementation points are refined in combination with the execution requirements of the process task link, supplemented with adjustment details adapted to the process task link, and the adjustment details include text interaction content specification, text interaction timing requirements, and text interaction feedback standards.

[0099] For the matched adjustment direction and implementation points, the execution requirements of the process task link are refined. For example, the process task link is the development link, and its execution requirements include promoting development according to module division, submitting daily development progress, and passing static code check. When refining the adjustment direction of "standardizing text expression format", the supplemented text interaction content specification is "the development link text needs to indicate the module name, development progress percentage, and whether it has passed local debugging"; the text interaction timing requirement is "daily development progress feedback needs to be submitted before 17:00"; and the text interaction feedback standard is "the feedback content needs to include the number of defects and repair plan".

[0100] Step S1531: The execution requirements of the process task link are retrieved, which include the task input standard, task processing flow, and task output requirement of the process task link.

[0101] The execution requirements of the corresponding process task link are retrieved from the process task management document. Taking the development link as an example, the task input standard is that the design document has passed the review and the development environment has been completed; the task processing flow is to allocate tasks according to modules, code writing, local debugging, and code submission; and the task output requirement is to complete module code writing, submit to the code repository, and pass static code check.

[0102] Step S1532: Analyze the adaptability of the interaction rule adjustment direction and the execution requirements, and determine the part of the adjustment direction that needs to be combined with the execution requirements.

[0103] The adaptability of the interaction rule adjustment direction and the execution requirements is analyzed to find out the content in the adjustment direction that is associated with the execution requirements. For example, the adjustment direction "specifying the text expression format" needs to be combined with the task output requirement in the execution requirements, because the output requirement involves code submission and inspection results, which need to be accurately expressed in text interaction; at the same time, it is related to the daily progress submission link in the task processing flow, and the expression format needs to be clearly stated to adapt to the progress submission requirement.

[0104] Step S1533: For the part that needs to be combined, refer to the task input standard in the execution requirements to set the text interaction content specification; refer to the task processing flow in the execution requirements to set the text interaction timing requirement, and refer to the task output requirement in the execution requirements to set the text interaction feedback standard.

[0105] According to the adaptability analysis result, refer to the execution requirements to set specific adjustment details. Referring to the task input standard in the development link, the text interaction content specification needs to include "design document review status" and "development environment status". Referring to the daily progress submission link in the task processing flow, the text interaction timing requirement is set to "submit the daily development progress before 17:00". Referring to the task output requirement, the text interaction feedback standard is set to "the feedback content must include code submission status, static code inspection result and defect quantity".

[0106] Step S1534: Set the text interaction content specification, text interaction timing requirement and text interaction feedback standard as adjustment details adapted to the process task link.

[0107] The above set text interaction content specification, timing requirement and feedback standard are integrated into adjustment details adapted to the process task link. These details are specific supplements to the interaction rule adjustment direction and implementation points, ensuring that the adjustment suggestion can meet the actual process task execution requirements.

[0108] Step S154: Integrate the refined interaction rule adjustment direction, adjustment implementation points and adjustment details into node interaction rule adjustment suggestions.

[0109] The refined interaction rule adjustment direction, adjustment implementation points and adjustment details are integrated to form the system node interaction rule adjustment suggestion. For example, for the trigger factor of unclear text expression in the development link, the integrated adjustment suggestion is: adjustment direction "standardize text expression format"; implementation points "clear text content elements, use standardized terminology"; adjustment details "content specification needs to include module name and other elements, feedback needs to include code checking results before 17:00 every day".

[0110] Step S155: Extract the abnormal description information corresponding to the trigger factor and the link identification information of the process task link. The abnormal description information includes abnormal manifestation and abnormal influence range. The link identification information includes link name and link execution order.

[0111] The abnormal description information corresponding to the trigger factor is extracted from the abnormal record. The abnormal manifestation of unclear text expression is "text content does not clearly indicate task status and time node", and the abnormal influence range is "test node cannot accurately arrange test plan after receiving information". The link identification information of the process task link is extracted. The link name of the development link is "development link", and the link execution order is the 3rd (after the requirement analysis and design link, and before the test link).

[0112] Step S156: Integrate the node interaction rule adjustment suggestion, abnormal description information and link identification information into the core content of the process node optimization instruction.

[0113] The node interaction rule adjustment suggestion, abnormal description information and link identification information are organically integrated to form the core content of the process node optimization instruction. The core content needs to clearly present the abnormal situation, adjustment suggestion and the link it belongs to, for example, "the development link (execution order 3rd) has unclear text expression abnormality, which is manifested as the text does not clearly indicate the task status and time node, affecting the test plan arrangement. It is suggested to standardize the text expression format, clearly indicate the content elements, and submit the feedback including code checking results before 17:00 every day".

[0114] Step S157: Add instruction number, generation time and target process control terminal identification to the core content to form a complete process node optimization instruction containing node interaction rule adjustment suggestion.

[0115] A unique instruction number is added to the core content, such as ZC-20250821-001, the specific time of instruction generation is recorded, such as August 21, 2025 10:30:00, and the target process control terminal identification is specified, such as KZ-001 (development node control terminal). After adding these information, a complete process node optimization instruction is formed, ensuring the uniqueness, traceability and accurate delivery of the instruction.

[0116] Step S158: Send the process node optimization instruction to the process control terminal.

[0117] The complete process node optimization instruction is sent to the target process control terminal through the network communication module. After the process control terminal receives the instruction, it modifies the node interaction rules according to the adjustment suggestions in the instruction, optimizes the text interaction of the process node, and reduces the occurrence of subsequent exceptions.

[0118] Step S210: Collect multiple sets of process node text monitoring sample data, each set of sample data containing process node text interaction and task phase corresponding relationship sample, task phase matching degree sample, and corresponding text interaction exception tracing label.

[0119] In the model pre-training stage, a large number of process node text monitoring sample data is first collected. These sample data come from historical software development project process monitoring records, and each set of sample data contains text interaction and task phase corresponding relationship sample, such as "J002 development node-module development phase-text A"; task phase matching degree sample, i.e. the matching degree value of the corresponding relationship; and corresponding text interaction exception tracing label, which clearly indicates whether the sample has an exception, the type of exception, and the exception path, such as "there is a phase misplacement exception, path J002-J003".

[0120] Step S220: Perform structured conversion processing on the corresponding relationship sample to convert unstructured corresponding relationship into structured relationship data table, which contains text interaction identifier, task phase identifier, and node identifier.

[0121] Unstructured corresponding relationship samples, such as natural language description "development node sends text X in module development phase", are converted into structured relationship data table. The column fields of the table are text interaction identifier (such as WJ-001), task phase identifier (such as JD-002), and node identifier (such as J002), and each row records the structured information of a set of corresponding relationship samples, facilitating subsequent vector conversion processing.

[0122] Step S230: Perform feature normalization processing on the task phase matching degree sample to obtain the normalized task phase matching degree sample.

[0123] The numerical range of the task phase matching degree sample may vary, so feature normalization processing is performed to ensure the stability of model training. The min-max normalization method is used to map the matching degree sample value to the interval of 0-1, and the calculation formula is normalized value=(original value-min value) / (max value-min value), where min value and max value are the minimum and maximum values of all matching degree samples. The normalized task phase matching degree sample is obtained after processing.

[0124] Step S240: converting the relational data table into a vector form to obtain a relational feature sample vector, and converting the normalized task stage matching degree sample into a vector form to obtain a matching degree feature sample vector.

[0125] The text interaction identifier, the task stage identifier and the node identifier in the relational data table are subjected to numerical coding processing using a coding tool. For the text interaction identifier, a unique integer coding mode is adopted, and each different text interaction identifier corresponds to a unique integer value, for example, the text interaction identifier W001 is coded as 1, the text interaction identifier W002 is coded as 2, and so on. The task stage identifier is also subjected to integer coding, for example, the requirement collection stage identifier G001 is coded as 101, and the module development stage identifier G002 is coded as 102. The coding rule of the node identifier is consistent with the above two, the requirement analysis node J001 is coded as 201, and the development node J002 is coded as 202.

[0126] The coded text interaction identifier, the task stage identifier and the node identifier are combined in order to form a feature array of each record in the relational data table. For example, the text interaction identifier of a certain record is coded as 5, the task stage identifier is coded as 102, and the node identifier is coded as 202, and the corresponding feature array is [5, 102, 202]. The length of the feature array of all records is uniformly processed, and if there are arrays of different lengths, they are adjusted to the same length by zero padding to ensure the consistency of subsequent vector conversion.

[0127] After the feature array processing is completed, it is converted into a relational feature sample vector. The dimension of each relational feature sample vector is consistent with the length of the feature array, and each element in the vector corresponds to a coding value in the feature array. For example, the feature array [5, 102, 202] is converted into a relational feature sample vector, and the three dimensions of the vector are 5, 102 and 202 respectively.

[0128] For the normalized task stage matching degree sample, since it is in the form of a single numerical value, the numerical value of each sample is directly taken as the only element of the matching degree feature sample vector. For example, the normalized task stage matching degree sample value is 0.85, and the corresponding matching degree feature sample vector is [0.85]. Through the above method, different types of sample data are converted into a vector form that can be processed by the model.

[0129] Step S250: concatenating the relational feature sample vector and the matching degree feature sample vector to generate a sample feature vector for each group of sample data.

[0130] After obtaining the relation feature sample vector and the matching degree feature sample vector, they are combined according to the feature concatenation rules. The concatenation operation is performed in the order of relation feature sample vector first, then matching degree feature sample vector, arranging the elements of the two vectors sequentially to form a new sample feature vector. For example, if the relation feature sample vector of a certain set of sample data is [5,102,202] and the matching degree feature sample vector is [0.85], the sample feature vector generated after concatenation is [5,102,202,0.85].

[0131] For relational feature sample vectors of different lengths, since the length has been standardized by padding with zeros in previous processing, the length of the concatenated sample feature vectors remains consistent. If the standardized length of the relational feature sample vectors is 10 and the length of the matching degree feature sample vectors is 1, then the length of the sample feature vector for each group of sample data is 11. Through the above concatenation method, the sample feature vectors simultaneously contain the association information between text interaction and task stages, as well as the matching degree information.

[0132] Step S260: Divide the sample feature vector and the corresponding text interaction anomaly tracing label into a training set, a validation set and a test set.

[0133] The sample feature vectors and their corresponding text interaction anomaly tracing labels are grouped according to a preset division ratio. In this embodiment, the training set, validation set, and test set are divided in a 7:2:1 ratio. First, all sample data are randomly shuffled to ensure the randomness of the sample distribution and avoid training bias caused by the order of samples.

[0134] Assuming a total of 10,000 samples, after random shuffling, the first 7,000 samples are used as the training set for model parameter learning; the middle 2,000 samples serve as the validation set for model performance evaluation during training; and the last 1,000 samples serve as the test set for verifying the model's generalization ability after training. The feature vector of each sample corresponds one-to-one with its corresponding text interaction anomaly tracing label, stored in the feature and label files of the three datasets respectively, for easy retrieval during model training.

[0135] Step S270: Construct the network structure of the process node text monitoring model. The network structure of the process node text monitoring model includes a feature association module and an anomaly tracing module. The feature association module of the process node text monitoring model includes a relation encoding submodule and a matching degree conversion submodule. The anomaly tracing module of the process node text monitoring model includes multiple parallel feature extraction submodules.

[0136] The process node text monitoring model adopts a deep learning network architecture, which is divided into two main parts: a feature association module and an anomaly tracing module. The feature association module is responsible for processing and fusing the input features, while the anomaly tracing module performs anomaly identification and tracing based on the fused features.

[0137] The relation encoding submodule in the feature association module employs a combination of embedding and fully connected layers. The embedding layer maps the dimensions of the input relation feature vector, transforming high-dimensional, sparse encoded features into low-dimensional, dense feature representations. The fully connected layer performs a non-linear transformation on the embedded features, enhancing their expressive power through activation functions. The matching degree transformation submodule is a single fully connected layer that adjusts the dimensions of the input matching degree feature vector to match the output dimension of the relation encoding submodule, preparing for subsequent feature fusion.

[0138] The anomaly tracing module comprises multiple parallel feature extraction sub-modules, each corresponding to a specific type of text interaction anomaly. Each feature extraction sub-module consists of multiple convolutional and pooling layers. The convolutional layers extract local anomaly patterns from the features using kernels of different sizes, while the pooling layers downsample the convolutional results, preserving key features and reducing the number of parameters. The outputs of multiple feature extraction sub-modules are then combined through a concatenation layer and input into a fully connected classification layer, ultimately outputting the text interaction anomaly tracing result.

[0139] Step S280: Train the constructed process node text monitoring model using the training set, and calculate the loss value between the source tracing result output by the process node text monitoring model and the source tracing label of text interaction anomaly using the cross-entropy loss function.

[0140] The feature vectors of the training set samples are input into the constructed process node text monitoring model. The model outputs preliminary source tracing results after forward propagation. At the same time, the corresponding text interaction anomaly source tracing tags are retrieved from the training set. These tags contain pre-labeled anomaly types and anomaly path information.

[0141] The cross-entropy loss function is used to calculate the difference between the model's output source tracing result and the label. The magnitude of the loss value reflects the degree of deviation between the model's prediction and the actual situation. The cross-entropy loss function calculates the loss by measuring the distance between two probability distributions. The closer the probability distribution of the model's output source tracing result is to the probability distribution of the label, the smaller the loss value; conversely, the larger the loss value is. During training, the loss value of each batch of samples is recorded in real time.

[0142] Step S290: Adjust the network parameters of the process node text monitoring model according to the loss value using the backpropagation algorithm, including the weights of the relation encoding submodule of the process node text monitoring model, the coefficients of the matching degree transformation submodule of the process node text monitoring model, and the parameters of the feature extraction submodule of the process node text monitoring model.

[0143] After obtaining the loss value, the backpropagation algorithm is used to calculate the gradient of the loss value with respect to the network parameters of each layer along the network structure. For the weight parameters of the relation encoding submodule, the gradient magnitude and direction are adjusted to make the outputs of the embedding layer and the fully connected layer more closely match the true distribution of features. The coefficient adjustment of the matching degree transformation submodule focuses on optimizing the fit between the matching degree features and the relation features, ensuring the feature complementarity of the two during fusion.

[0144] In the anomaly tracing module, the convolutional kernel parameters and bias terms of the feature extraction submodule are updated using gradient descent to adjust the weight distribution of the convolutional kernels and enhance their ability to capture anomalous features. The weights and biases of the fully connected layers are also adjusted based on the gradient to optimize the model's classification accuracy for anomaly types. After each parameter adjustment, the model's loss value gradually decreases, and the prediction accuracy gradually improves.

[0145] Step S2100: After each preset training round, the performance of the process node text monitoring model is evaluated using the validation set, and the anomaly tracing accuracy of the process node text monitoring model on the validation set is calculated.

[0146] The preset training rounds are set to 10 rounds, meaning that after every 10 rounds of training, the model's performance is evaluated using a validation set. The feature vectors of the validation set samples are input into the model in its current training state to obtain the model's output source attribution results. These source attribution results are compared with the text interaction anomaly attribution labels in the validation set, and the proportion of correctly predicted samples out of the total number of samples is calculated; this proportion is the anomaly attribution accuracy.

[0147] For example, if the validation set contains 2000 samples, and the model correctly predicts 1800 samples, then the anomaly tracing accuracy is 1800 / 2000 = 0.9, or 90%. Simultaneously, recall and precision for different anomaly types can be calculated to comprehensively evaluate the model's recognition performance across various anomalies.

[0148] Step S2110: If the accuracy of the anomaly tracing reaches the preset accuracy standard, stop training the process node text monitoring model; if it does not reach the standard, adjust the learning rate of the process node text monitoring model and continue training until the accuracy standard is reached.

[0149] The preset accuracy standard is set to 92%. When the anomaly tracing accuracy on the validation set reaches or exceeds 92%, it indicates that the model has reached the expected performance requirements. The training process is then stopped, and the network parameters of the current model are saved as the final pre-trained model parameters.

[0150] If the accuracy does not meet the standard, the model's learning rate needs to be adjusted. The initial learning rate is set to 0.001. When the accuracy fails to meet the standard, the learning rate is reduced to 0.5 times the original value, i.e., adjusted to 0.0005, and training continues. After each adjustment of the learning rate, a new round of training and validation is performed until the anomaly identification accuracy reaches the preset standard. Adjusting the learning rate helps prevent the model from converging too slowly or overfitting during training.

[0151] Step S2120: Use the test set to test the generalization ability of the trained process node text monitoring model, and verify the traceability effect of the process node text monitoring model on unseen sample data.

[0152] The feature vectors of the test set samples are input into the trained model, and the model outputs the corresponding text interaction anomaly attribution results. Since the sample data in the test set has not been used during model training, it can effectively verify the model's adaptability to new data. The attribution results of the test set are compared with the corresponding labels, and the anomaly attribution accuracy, recall, and F1 score on the test set are calculated.

[0153] If the metrics on the test set are close to those on the validation set and all meet the preset standards, it indicates that the model has good generalization ability. If the metrics differ significantly, it is necessary to re-examine the model structure or training process, and if necessary, increase the number of training samples or adjust the network parameters until the model's generalization ability meets the requirements.

[0154] Step S2130: The process node text monitoring model that passes the generalization ability test is identified as the pre-trained process node text monitoring model.

[0155] After validating the model's generalization ability on the test set, if all performance metrics meet expectations, the model is designated as the pre-trained process node text monitoring model. The model's network structure, parameter configuration, and performance metrics are recorded in detail and stored in the model repository. Simultaneously, a user manual is generated, clearly defining the model's input and output formats, applicable scenarios, and invocation methods. The pre-trained model can be directly used to handle anomalies in new process node text interaction data, enabling real-time monitoring and anomaly alerts for process nodes.

[0156] Furthermore, Figure 2A schematic diagram of the hardware structure of a process node text monitoring system 100 based on AI natural language processing for implementing the methods provided in the embodiments of this application is shown. Figure 2 As shown, the AI ​​natural language processing-based process node text monitoring system 100 may include at least one processor 102 (the processor 102 may be, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, a transmission device 106 for communication functions, and a controller 108. Those skilled in the art will understand that... Figure 2 The structure shown is for illustrative purposes only and does not limit the structure of the AI-based natural language processing-based process node text monitoring system 100. For example, the AI-based natural language processing-based process node text monitoring system 100 may also include more than Figure 2 The more or fewer components shown, or having the same Figure 2 The different configurations shown.

[0157] The memory 104 can be used to store software programs and modules of the application software, such as the program instructions corresponding to the method embodiments described above in this application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the above-described method for monitoring process nodes based on AI natural language processing. The transmission device 106 is used to acquire or send data via a network.

[0158] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

Claims

1. A method for monitoring process node text based on AI natural language processing, characterized in that, The method includes: Capture text interaction data and corresponding process progress information generated by each process node during process execution. The text interaction data includes the text content sent and received by each process node, and the process progress information includes the current task stage being executed by each process node. The text interaction data and the process progress information are correlated and modeled to obtain the correspondence between the text interaction and the task stage of each process node and the matching degree of the task stage corresponding to the text interaction. The pre-trained process node text monitoring model is invoked to perform anomaly tracing on the correspondence and the matching degree of the task stage, and generate anomaly tracing results for text interaction of each process node. Based on the results of the text interaction anomaly tracing, determine the triggering factors of the text interaction anomalies in each process node and the process task links corresponding to the abnormal text interactions. Based on the triggering factors and the process task steps, a process node optimization instruction containing suggestions for adjusting node interaction rules is generated, and the process node optimization instruction is sent to the process control terminal.

2. The process node text monitoring method based on AI natural language processing according to claim 1, characterized in that, The step of performing correlation modeling on the text interaction data and the process progress information to obtain the correspondence between text interaction and task stage at each process node and the matching degree of the task stage corresponding to the text interaction includes: Semantic parsing is performed on the text content in the text interaction data to extract semantic elements that represent the task execution status. The semantic elements include a description of the task execution progress, a description of the task requirements, and feedback on the task results. The task stages in the process progress information are characterized by feature identification, and a unique stage feature identifier is set for each task stage. The stage feature identifier includes the task stage name, task stage execution requirements and task stage output standards. The extracted semantic elements are semantically matched with the stage feature identifiers of each task stage to determine the task stage corresponding to each text interaction data, thus forming a correspondence between the text interaction of each process node and the task stage. Calculate the semantic similarity between the semantic elements of each text interaction data and the stage feature identifier of the corresponding task stage, and use the semantic similarity as the matching degree of the task stage corresponding to the text interaction. The correspondence and the task stage matching degree are bound to node identifiers so that each correspondence and task stage matching degree corresponds to the identifier of the process node that generates text interaction data, thereby obtaining the correspondence between text interaction and task stage of each process node and the task stage matching degree corresponding to the text interaction.

3. The process node text monitoring method based on AI natural language processing according to claim 2, characterized in that, The step of performing semantic parsing processing on the text content in the text interaction data to extract semantic elements representing the task execution state from the text content includes: The text content in the text interaction data is divided into multiple sentence units according to the sentence segmentation rules; The pre-trained semantic understanding model is invoked to perform semantic representation transformation on each sentence unit, converting each sentence unit into a sentence semantic vector that can represent its semantic connotation; Retrieve a preset task status semantic dictionary, which contains feature words representing the task execution status and corresponding semantic vectors. The feature words include task in execution, task pending confirmation, task completed, and task needs adjustment. Calculate the similarity between the semantic vector of each statement unit and the semantic vector of each feature word in the task state semantic dictionary; The semantic information associated with the semantic vectors of the feature words with the highest similarity is selected, and this semantic information is determined as the semantic element representing the task execution state in the sentence unit; By integrating the semantic elements corresponding to all sentence units in each text content, we obtain the semantic elements in the text content that represent the task execution status.

4. The process node text monitoring method based on AI natural language processing according to claim 1, characterized in that, The pre-trained process node text monitoring model is invoked to perform anomaly tracing processing on the correspondence and the task stage matching degree, generating text interaction anomaly tracing results for each process node, including: The correspondence and the task stage matching degree are input into the feature association module of the process node text monitoring model. The correspondence is structured and encoded to obtain a relationship feature vector. The task stage matching degree is numerically transformed to obtain a matching degree feature vector. In the feature association module of the process node text monitoring model, the relationship feature vector and the matching degree feature vector are subjected to feature fusion processing to generate a fused feature vector. The feature fusion processing is performed through feature connection processing so that the fused feature vector contains both corresponding relationship information and matching degree information. The fused feature vector is input into the anomaly tracing module of the process node text monitoring model. The anomaly tracing module of the process node text monitoring model contains multiple parallel tracing sub-modules, each of which corresponds to a text interaction anomaly type. Each source tracing submodule performs anomaly feature extraction processing on the fused feature vector, and identifies feature elements in the fused feature vector that are related to the corresponding anomaly type. The feature elements include text interaction semantic features corresponding to the anomaly and task stage features corresponding to the anomaly. The path of the anomaly is traced based on the extracted feature elements, including the sending node, receiving node, and transmission order of the abnormal text interaction in the process. Integrate the abnormal feature elements and abnormal propagation paths output by each tracing submodule to generate text-based interactive abnormal tracing results for each process node.

5. The process node text monitoring method based on AI natural language processing according to claim 4, characterized in that, Each source tracing submodule performs anomaly feature extraction processing on the fused feature vector, identifying feature elements in the fused feature vector that are related to the corresponding anomaly type, including: In each source tracing submodule, the anomaly feature template associated with the corresponding anomaly type is retrieved. The anomaly feature template is a typical feature distribution form of the fused feature vector under that anomaly type. The fused feature vector and the abnormal feature template are subjected to feature comparison processing to identify feature regions in the fused feature vector that are consistent with the feature distribution of the abnormal feature template. Extract feature elements within the feature region, including the numerical value of the feature element, the position of the feature element in the fused feature vector, and the text interaction and task stage association information corresponding to the feature element; The extracted feature elements are subjected to semantic association verification processing to confirm the semantic association between the text interaction semantics corresponding to the feature elements and the semantic association between the anomaly type. The semantic association is judged by the similarity between the semantic vector corresponding to the feature element and the semantic vector of the anomaly type. Feature elements whose semantic relevance meets the preset requirements are identified as feature elements related to the corresponding exception type.

6. The process node text monitoring method based on AI natural language processing according to claim 1, characterized in that, The step of determining the triggering factors of text interaction anomalies in each process node and the corresponding process task links of the abnormal text interactions based on the text interaction anomaly tracing results includes: The abnormal feature elements in the text interaction anomaly tracing results are analyzed, and the semantic information representing the abnormal cause in the abnormal feature elements is identified. The semantic information includes text expression deviation, task stage cognitive deviation, and interaction rule understanding deviation. Retrieve historical text interaction data for each process node and analyze the text interaction patterns of each process node before the anomaly occurs. The text interaction patterns include text interaction frequency, text interaction content type, and the distribution of task stages corresponding to the text interaction. Compare the text interaction data at the time of the anomaly with the text interaction pattern to identify the differences between the two. The differences include changes in text interaction frequency, changes in text interaction content type, and shifts in the task stage corresponding to the text interaction. Based on the semantic information and the differences, the triggering factors for text interaction anomalies are determined. The triggering factors include unclear text descriptions, cognitive errors at the task stage, and deviations in the execution of interaction rules. Retrieve the process task segmentation information, determine the process task segment to which the abnormal text interaction belongs based on the task stage corresponding to the abnormal text interaction, and obtain the process task segment corresponding to the abnormal text interaction.

7. The process node text monitoring method based on AI natural language processing according to claim 6, characterized in that, The process of retrieving historical text interaction data from each process node and analyzing the text interaction patterns of each process node before the anomaly occurred includes: Obtain the text interaction history data of each process node within a preset time period before the anomaly occurs. The text interaction history data includes all text content sent and received by each process node within that time period and the corresponding task stage. The text content in the text interaction history data is classified and processed according to the semantic type of the text content into task instruction type, task feedback type and task consultation type. The number of times various types of text content appear in different task stages is counted, and the proportion of various types of text content in each task stage is calculated to obtain the distribution of text interaction content types in each task stage. The total number of times text content is sent and received by each process node within a preset time interval is counted, the number of text interactions per unit time is calculated, and the text interaction frequency of each process node is obtained. By integrating the distribution and the frequency of text interactions, a pattern of text interactions at each process node before an anomaly occurs is formed.

8. The process node text monitoring method based on AI natural language processing according to claim 1, characterized in that, The process node optimization instruction, which includes suggestions for adjusting node interaction rules, is generated based on the triggering factors and the process task stages, including: Retrieve a preset abnormal triggering factor response strategy library, which includes the interaction rule adjustment direction and adjustment implementation points for each triggering factor; Based on the aforementioned triggering factors, the corresponding interaction rules are matched from the abnormal triggering factor response strategy library to adjust the direction and key points of adjustment implementation; Based on the execution requirements of the process task stages, the adjustment direction and key points of the matched interaction rules are refined, and adjustment details adapted to the process task stages are supplemented. The adjustment details include text interaction content specifications, text interaction timing requirements, and text interaction feedback standards. The refined interaction rules adjustment direction, key points of adjustment implementation, and adjustment details are integrated into node interaction rule adjustment suggestions; Extract the abnormal description information corresponding to the triggering factor and the step identification information of the process task. The abnormal description information includes the abnormal manifestation and the scope of the abnormal impact. The step identification information includes the step name and the step execution order. The node interaction rule adjustment suggestions, anomaly description information, and process identification information are integrated into the core content of the process node optimization instruction; Add instruction number, generation time and target process control terminal identifier to the core content to form a complete process node optimization instruction that includes suggestions for adjusting node interaction rules; Specifically, the adjustment direction and key points of the matched interaction rules are refined based on the execution requirements of the process task stages, supplementing adjustment details to adapt to the process task stages, including: Retrieve the execution requirements of the process task step, the execution requirements including the task input standards, task processing flow and task output requirements of the process task step; Analyze the compatibility between the adjustment direction of the interaction rules and the execution requirements, and determine the part of the adjustment direction that needs to be combined with the execution requirements; For the parts that need to be combined, the text interaction content specifications are set with reference to the task input standards in the execution requirements; the text interaction timing requirements are set with reference to the task processing flow in the execution requirements; and the text interaction feedback standards are set with reference to the task output requirements in the execution requirements. The established text interaction content specifications, text interaction timing requirements, and text interaction feedback standards will be used as adjustment details to adapt to the process task stages.

9. The process node text monitoring method based on AI natural language processing according to claim 1, characterized in that, The pre-training process of the process node text monitoring model includes: Collect multiple sets of text monitoring sample data for process nodes. Each set of sample data includes a sample of the correspondence between text interaction and task stage of the process node, a sample of the matching degree of task stage, and a corresponding text interaction anomaly tracing tag. The correspondence samples are subjected to a structured transformation process to convert the unstructured correspondences into a structured relational data table, which includes text interaction identifiers, task stage identifiers, and node identifiers. The task stage matching degree samples are subjected to feature normalization processing to obtain normalized task stage matching degree samples. The relational data table is converted into vector form to obtain relational feature sample vectors, and the normalized task stage matching degree samples are converted into vector form to obtain matching degree feature sample vectors. The relation feature sample vector and the matching degree feature sample vector are concatenated to generate a sample feature vector for each group of sample data. The sample feature vectors and their corresponding text interaction anomaly tracing labels are divided into training set, validation set and test set; A network structure for a process node text monitoring model is constructed. The network structure of the process node text monitoring model includes a feature association module and an anomaly tracing module. The feature association module of the process node text monitoring model includes a relation encoding submodule and a matching degree conversion submodule. The anomaly tracing module of the process node text monitoring model includes multiple parallel feature extraction submodules. The constructed process node text monitoring model is trained using the training set, and the loss value between the source tracing result output by the process node text monitoring model and the source tracing label of text interaction anomaly is calculated using the cross-entropy loss function. The network parameters of the process node text monitoring model are adjusted based on the loss value using the backpropagation algorithm, including the weights of the relation encoding submodule, the coefficients of the matching degree transformation submodule, and the parameters of the feature extraction submodule. After each preset training round, the performance of the process node text monitoring model is evaluated using the validation set, and the anomaly tracing accuracy of the process node text monitoring model on the validation set is calculated. If the accuracy of the anomaly tracing reaches the preset accuracy standard, the training of the process node text monitoring model is stopped; if it does not reach the standard, the learning rate of the process node text monitoring model is adjusted and training continues until the accuracy standard is reached. The generalization ability of the trained process node text monitoring model was tested using the test set to verify the traceability effect of the process node text monitoring model on unseen sample data. The process node text monitoring model that passes the generalization ability test is identified as the pre-trained process node text monitoring model.

10. A process node text monitoring system based on AI natural language processing, characterized in that, The method includes a processor and a readable storage medium storing a program that, when executed by the processor, implements the process node text monitoring method based on AI natural language processing as described in any one of claims 1-9.

Images